A place where everything changes
AI in Education and Training
Whether we like it or not, AI has already entered higher education and the workplace. It meets woefully unprepared structures and people.
In schools and academic life, teachers lack understanding, kids embrace it wholeheartedly while at the same time being overwhelmed and scared by it. Grading isn't prepared by the challenges of telling true mastery from polished outputs. If not guided properly, AI will damage education for generations to come, and in some ways irrecoverably.
In companies, AI use is picking up, both authorized and in the dark. Some people embrace it, some reject it, few understand it sufficiently to know what it can and cannot do, and how it affects the way they work, produce relevant output - and how they become and stay experts.
Beyond the sometimes questionable quality of AI output, the key issue - in education and companies - is cognitive offloading, the delegation of steps at the core of a task to AI: understanding the objective, researching and analyzing the data, evaluating options and consequences, and finally turning all this into an argument and meaningful output. For students and job entrants, delegating these steps to AI mean that they will never truly develop these capabilities, let alone the ability to guide and judge AI actions. For experienced workers, this kind of delegation to AI demonstrably leads to deskilling, the loss of previously acquired abilities.
At the same time, we have to acknowledge that AI systems can help doing the job better and make us faster in reaching our goals, there is no denying it. Studies also show that AI-assisted learning can have benefits for skill acquisition and learning speed, without jeopardizing the actual learning outcome. Both require the right approach to AI.
This is why 9senses is creating a non-profit arm that provides support to individuals and institutions alike with the aim of making the inevitable arrival of AI in our world beneficial, and not detrimental.
La formation de base gratuite à l’IA
Teaching in a world with AI
AI Literacy for Students
“The downside of adult offloading is that people get less sharp. The downside of adolescents growing up delegating their minds to AI is a generation that was never sharp to begin with.”
Timothy Cook, M.Ed., Psychology Today, 2026
Below, we provide a first overview of how AI affects education systems and companies, and how we might be able to deal with it in a way that impact is positive instead of detrimental. Stay tuned for more as we build our content. Soon, each of the segments will also include their own education materials.
Different dynamics
The problem is the same everywhere. AI enters anyway. What you can actually do about it is not. See what it means for each group, plus what happens if we don't act.
Students and AI
How does AI affect the learning of skills schools and universities should build, and how could AI be used beneficially?
Teachers and AI
What is AI? How can teachers and students benefit from it? and how not to lose control as AI enters the learning environment.
Institutions and AI
What can institutions do to ensure AI doesn't jeopardize the learning outcome without endangering fairness?
Employees and AI
How to ensure seniors and juniors grow with and despite increasing AI use in your business? How to prevent deskilling?
First and foremost, AI doesn't have to be inherently bad. If it is used as a tutor that explains and answers questions, it can help adjust learning pace and format to each student's individual needs, like a one-on-one tutor could. This can support understanding and learning without damaging them, as a randomized trial with roughly a thousand maths students showed. Unrestricted GPT-4 access helped them solve tasks but tanked their scores when tested without AI help available, whereas a hints-only AI tutor helped in task resolution and didn't negatively affect test scores (Bastani et al., 2025).
But unfortunately, most students don't use AI that way when given the choice. An Anthropic analysis of more than half a million student requests showed that a majority were using AI for producing output (Anthropic, 2025). And when they do, they skip the exact purpose what makes them learn: understanding a question, searching for evidence, selecting and structuring relevant material, turning their thoughts into a solid argument and finally producing coherent output understandable for the recipients. By delegating most of that to AI, the gradable output is there, but the learning isn't, which studies link to measurably weaker critical thinking abilities (Gerlich, 2025). MIT’s EEG study calls the result cognitive debt (Kosmyna et al., 2025).
While using AI, self-confidence erodes. How can you claim a result as yours if it was mostly generated by AI, and you simply read and memorized the output? That a fluent answer feels understood until it must be explained unaided was described long before AI (Rozenblit and Keil, 2002), and MIT’s students often could not quote essays they had just submitted. So how should you ever be confident about your abilities, if nothing was yours to begin with?
And then there is the problem of evaluation. If grading cannot reliably tell the product of hard work from AI-generated output, scores become irrelevant and erode students' trust in the system even further. If three weeks of hard work give a brilliant student 14 of 15 points, while those who produce output with AI and disguise their cheating well (Perkins et al., 2024) arrive at 15, everything is off. And it becomes worse once detectors are incorrectly accusing brilliant students and non-native speakers (Liang et al., 2023) of cheating with AI.
All this needs fixing. And fast.
References
Gerlich, M. (2025). AI tools in society: impacts on cognitive offloading and the future of critical thinking. Societies, 15(1), 6.
Kosmyna, N. et al. (2025). Your brain on ChatGPT: accumulation of cognitive debt when using an AI assistant for essay writing task. MIT Media Lab, arXiv:2506.08872.
Rozenblit, L. and Keil, F. (2002). The misunderstood limits of folk science: an illusion of explanatory depth. Cognitive Science, 26(5).
Liang, W. et al. (2023). GPT detectors are biased against non-native English writers. Patterns, 4(7).
Perkins, M. et al. (2024). Simple techniques to bypass GenAI text detectors: implications for inclusive education. International Journal of Educational Technology in Higher Education, 21.
Bastani, H. et al. (2025). Generative AI without guardrails can harm learning: evidence from high school mathematics. PNAS, 122(26).
Anthropic (2025). Anthropic Education Report: how university students use Claude.
Anthropic (2026). How AI assistance impacts the formation of coding skills.
What doing nothing does
If we don't act swiftly, we will be producing a lost generation, a cohort of people who have never learned to use their mind and will be even less well equipped to direct AI and evaluate its output.
AI doesn't have to be the enemy of teaching - used well, it can do things a teacher rarely has time to do. For example, AI can explain the same concept in five different ways without getting tired for five students with different abilities and learning styles. Or it can help prepare high-quality teaching materials that can be adapted quickly to changes in the environment or current developments that make them more interesting to students.
But that's the theory. The downsides are there too. A paper currently under review found that when teachers used AI to generate lecture notes, assignments and exams, students rated the classes less enjoyable, less interesting and less important than the control group developing materials conventionally - this was observed particularly for teachers who used AI materials as they came, without using AI rather as input-providers or support (Sungu et al., 2026, under review). The rationale behind it is what is a general observation made in AI use, a detachment of the actual owner - here the teacher - from the content and the format of the output. Ownership and passion are related to the amount of thinking and work that went into the materials.
What AI also cannot do is policing: Ai-based detectors of AI use in work are notoriously unreliable and can easily be tricked, and they are more prone of flagging work of excellent students and those with non-native language backgrounds. This led to many universities turning the detectors off. Teaching with AI present has to assume we cannot tell the difference.
This moves the process of grading to the process, and to what can be demonstrated in person, without AI present. Not grading the polished paper or the glossy presentation, but rather the documentation of the progress to the result, having a 10-minute oral discussion on the subject, this immediately exposes AI use and rewards the real work. But it means shifting away from grading final grammar, structure and layout - the things AI now supplies for free. Fluency of presented output simply has stopped being a signal, which lets move many universities away from these formats.
At the same time, we have to teach the next generation how to deal with AI, because it is likely here to stay. In a way that explains it advantages and disadvantages openly, without glorification or condemnation. If we manage to do this in the classroom, we are no longer victims of Ai.
References
Bastani, H. et al. (2025). Generative AI without guardrails can harm learning: evidence from high school mathematics. PNAS, 122(26).
Sungu, A. et al. (2026). Generative AI can harm teaching. Working paper, under review; reported by The Hechinger Report.
Vanderbilt University (2023). Guidance on AI detection and why we’re disabling Turnitin’s AI detector.
Liang, W. et al. (2023). GPT detectors are biased against non-native English writers. Patterns, 4(7).
Perkins, M. et al. (2024). Simple techniques to bypass GenAI text detectors: implications for inclusive education. International Journal of Educational Technology in Higher Education, 21.
University of Queensland, in TEQSA (2024). Principles, criteria and standards for assessment with gen AI use.
University of Sydney (2024). Interactive oral assessment in practice.
Anthropic (2026). How AI assistance impacts the formation of coding skills.
Further reading: Northeastern University (2025) task-labelling standards; University of Bath assessment categorisation; the AI Assessment Scale; UNESCO AI Competency Framework for Teachers; Times Higher Education (2024) on sampled vivas.
What doing nothing does
The first thing you lose is the signal. A grade that can no longer separate a worked argument from a generated one has stopped measuring what you are trying to develop, while still deciding your students' futures.
The second thing you lose is the honest student. They watch the shortcut score higher and draw the obvious conclusion. Nothing you say in class outweighs what your marking rewards.
The third is quieter. Left untaught, students find the delegation mode by themselves, because it is the path of least resistance, and that is the mode that scored 24–39%. The cohort you hand on will be fluent, confident and unable to tell when the machine is wrong.
And you carry a risk you did not choose. Accusations built on detector output are being overturned on appeal, and they land disproportionately on international students.
A degree used to be a certificate of capability, but this promise is currently eroding. And schools and universities alike will have to realign significant parts of their teaching and evaluation mechanisms to remain in that position as AI creeps into every part of their curriculum. And while they defend that promise of teaching what is important, they also have to teach students how to deal with the unavoidable, AI in all its forms.
The most urgent part is the evidence standards. It is neither acceptable to wave shiny results through that have been generated, nor is it ok to accuse students of AI use if there is no certainty. And there will be no certainty if a final results, as AI, detectors and disguising efforts will play tag eternally - what will be a certain method of detecting AI use today will be outdated tomorrow, even the watermark concept recently announce by Anthropic will find a countermeasure.
However, there are easy ways to still keep evaluation honest, but they are costly. Shifting back to closed-door assignments, oral exams, and a thorough evaluation of documented steps that led to a final product, these all will prevent the use of AI instead of a student's brainpower. But they require more effort compared to grading polished final results, possibly with the use of AI. Finding the right balance matters, and some universities are already moving ahead with a clear two-lane approach that combines assignments with allowed (but documented) AI use with traditional methods that measure unassisted performance.
And allowing the - correct - assisting use of AI is something that is also part of the teaching experience students can expect in the class of 2026/27. They need to be prepared for a professional environment where being able to tell AI what to do, being able to evaluate its content, and to still own the final outcome, will be part of their job description.
References
Office of the Independent Adjudicator (2025). Casework note: complaints relating to AI and academic misconduct.
TEQSA (2024). Assessment reform for the age of artificial intelligence.
University of Sydney (2025). Frequently asked questions about the two-lane approach to assessment in the age of AI.
European Commission (2026). AI Omnibus enters into force.
EU AI Act: Article 4 (AI literacy), Article 5 (prohibited practices), Annex III point 3 (education and vocational training).
Institutions that do nothing will..
Your degree slowly stops being evidence of anything. It still certifies attendance and compliance, but not capability, which was the only thing anyone was buying.
Meanwhile you carry live legal exposure. Every misconduct case built on a detector score is a case you can lose, and losing them costs more than the cheating did. Vanderbilt did that arithmetic in 2023 and the arithmetic has not improved since.
The cost of reform only rises. Institutions that started early will have working models, trained staff and a defensible position while later movers are still forming a working group. Employers and rankings notice which group you are in, and they are already asking.
And the December 2027 deadline is not far away for anything that has to pass through procurement, a DPIA, an academic board and a senate.
The productivity is real and there is no case for refusing it. The difficulty is that the tasks being automated first are the ones a novice used to learn on — the research summary, the first draft, the routine analysis — so the output survives and the apprenticeship quietly does not.
It is measurable. In Anthropic’s study of skill formation, the group that delegated the work to AI learned significantly less and was not faster for it (Anthropic, 2026). BCG describes the organisational version as distributed de-skilling, with judgement, problem framing and evaluation of solutions the capabilities most exposed, and draws the conclusion that matters at board level: an organisation that has lost the ability to define the right problem cannot use AI to compensate (BCG, 2026).
The binding constraint is not headcount but mentor time. In a model from Central European University and Cambridge, senior capacity to absorb escalated problems is fixed whether or not juniors have AI, so teams get smaller and fewer juniors sit beside each senior — AI does not free Michelangelo’s time, because the bottleneck was never the carving (Fabbri et al., 2025). Learning and work were bundled together in apprenticeship; AI unbundles them, raising output per junior while the learning rate stays flat.
Nobody notices this happening to themselves, which is why it has to be measured with the tool switched off. A study of 126 airline pilots found that recency of hand-flying practice, not total hours, predicted who could still fly an approach manually; experience protected no one. Aviation’s answer was to train and assess manual proficiency deliberately, while warning against arbitrary quotas imposed without justification (FAA, 2017). The equivalent is periodic unassisted work judged against a person’s own earlier standard — a capability check, not surveillance.
Some firms are already buying the other option: IBM is tripling US entry-level hiring on the view that the companies doubling down now will be the ones that can still run themselves in five years (CIO, 2026). The capability you will need then is being built, or not built, in this year’s workflows — the full argument is in The lost generation.
References
Anthropic (2026). How AI assistance impacts the formation of coding skills.
BCG (2026). When everyone uses AI, companies risk critical skills.
Fabbri, F. et al. (2025). Broken ladders: AI and the apprenticeship model. Central European University and University of Cambridge.
FAA (2017). SAFO 17007: manual flight operations proficiency.
CIO (2026). IBM looks beyond short-term AI gains, tripling entry-level hiring.
Beane, M. (2024). How can we preserve human ability in the age of machines? MIT Sloan.
Nogueira dos Santos, J. et al. (2025). Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy. The Lancet Gastroenterology and Hepatology.
What doing nothing does
Margins improve first, which is why this is so easy to keep doing. You are drawing down a stock of expertise you have stopped replenishing, and the drawdown appears in no quarterly number.
Then the bench empties. When seniors retire or move on there is nobody to promote, and no external market to buy from either, because every competitor made the same saving in the same years. Scarcity of judgement is not a problem you solve with recruitment budget.
It is not only juniors. Nineteen experienced endoscopists, each with more than two thousand procedures behind them, saw their unassisted detection rate fall from 28.4% to 22.4% after a few months of AI-assisted work (The Lancet Gastroenterology and Hepatology, 2025). Experts losing a skill they spent years building, without noticing.
Doing nothing ends with an organisation running on AI-driven workflows that nobody inside it fully understands, and no capacity left to tell when they are quietly going wrong.
Questions employers ask us
Why hire and train juniors at all if AI is faster?
Because you are not buying this quarter's output, you are buying the people who will run the place in five years — and the productivity case is weaker than assumed: in the Anthropic study the AI-assisted group learned significantly less and was no faster. For a board, frame it as continuity risk rather than a training line. In three to five years the business will depend on AI-driven workflows, and the question is whether anyone inside it can still tell when those workflows are wrong. Same logic as R&D: the profits from not investing arrive first, and last least.
How do we know our seniors are not losing skill too?
You do not, unless you check, and they will not tell you because they cannot feel it. That is the finding from the endoscopy study and from aviation. Build in periodic unassisted work on real problems and compare against the person's own earlier standard rather than against a peer.
FAQs - Frequently Asked Questions
Can I just ban AI in my classroom?
You can ban it where you supervise, and that works because it is enforceable. A blanket ban on unsupervised work is not, and its main effect is to punish the students who respect it. Secure what carries the grade, and for everything else state which of your task labels applies and how use must be declared.
Do AI detectors work?
Not as evidence. Detection accuracy collapses against light editing and paraphrasing tools, and detectors misfire hardest on students writing in a second language, so a score is not something you can defend in an appeal — the UK's higher education ombudsman has been clear the burden of proof sits with the institution. Buying a better one is a subscription to an arms race you do not win; the same money buys more defensible ground spent on assessment design and panel training.
How do I grade fairly when I cannot tell who used AI?
Stop trying to tell and change what you grade. Move marks onto work you can observe: in-room writing, a short oral defence of a submitted piece, a viva where the student explains their own choices. Someone who did the thinking can defend it in five minutes and someone who did not, cannot, and you never had to accuse anyone of anything.
Isn't teaching AI use just teaching students to cheat better?
The opposite, on the evidence. Students who used AI to ask conceptual questions and check their understanding matched or beat the group with no AI; those using it unguided fell far behind. Nobody has to teach delegation. It is what students do when no one has shown them anything else.
Isn't securing assessment simply too expensive?
It is expensive, and pretending otherwise helps nobody: supervised and oral assessment is precisely what the sector moved away from to cope with student numbers. The point is that two-lane does not mean securing everything. Only the assessments that carry the qualification move into lane one, and that is a much smaller set than people assume once it is actually tabulated.